How do metadata and filtering improve retrieval?

Examples

• "What's the leave policy?" from an employee in India → filter country = IN, doc_type = policy, status = current.
• "What changed in v2.3 of the SDK?" → filter version = 2.3.
• Multi-tenant SaaS → always filter tenant_id = X (security, not just relevance).

Where metadata comes from

• Source systems (file path, author, modified date, space or folder).
• Parsing (section headings, page numbers).
• LLM extraction at index time (topics, entities, document type).

Self-querying retrieval: an LLM converts "Show me Q3 2026 sales reports from the APAC team" into {query: "sales report", filters: {quarter: "Q3-2026", team: "APAC"}}. Validate the extracted filters against allowed values, since the LLM can invent fields.

Pre-filter vs post-filter

• Pre-filter (filter, then search): correct results, but can be slow or reduce ANN recall with very selective filters.
• Post-filter (search, then filter): fast, but can return fewer than k results if most hits get filtered out.
• Modern vector DBs support filtered ANN that handles this in the engine.

Common mistakes

• Using metadata filters as the only security mechanism without testing them (see Q37).
• Over-filtering from wrongly extracted filters, which silently returns nothing.

Related

• Previous: Q9. How do you choose top-k, and what is the "lost in the middle" problem?
• Next: Q11. How do you prompt the LLM so it answers only from retrieved context and cites sources?
• LangChain
• LlamaIndex